Why this research matters
Speech carries affect through prosody, voice quality, timing, and lexical content, not only through the words that are spoken. Emotion recognition can support human–computer interaction, media analysis, accessibility, and wellbeing applications, but emotional expression varies across people, languages, and cultures. We study models that remain reliable beyond a single corpus, actor, or recording condition.
Research directions
- Acoustic and prosodic cues: Investigate how pitch, energy, timing, voice quality, and spectral structure contribute to perceived emotion.
- Self-supervised speech representations: Examine whether general-purpose speech models transfer to affective tasks and what information they capture or miss.
- Multimodal affect: Combine speech with text, visual, or contextual signals when these are available, while keeping the contribution of each modality interpretable.
- Cross-lingual and cross-corpus generalization: Study how models behave when language, speaker, domain, or recording conditions change.
- Uncertainty and calibration: Move beyond a single predicted label toward calibrated confidence, ambiguity, and subject-independent evaluation.
- Real-world conditions: Address noise, spontaneous speech, overlapping talk, and the difference between acted and natural emotional expression.
- Ethics and privacy: Consider consent, cultural variability, and the risk of overclaiming what a model can infer about a person’s inner state.
How we approach the work
We emphasize subject-independent evaluation, transparent preprocessing, and honest reporting of uncertainty. We treat emotion labels as annotations shaped by context and listener perception, not as direct measurements of a person’s internal experience.
What we aim to enable
Affect-aware systems that are useful and responsible: systems that can support people in realistic settings while making their limitations and assumptions clear.
